Every turn of an AI voice conversation pays for the same three components twice — speech-to-text, the language model, and text-to-speech — once in milliseconds and once in dollars. The milliseconds decide whether the agent feels conversational; the dollars decide whether it makes sense at volume. Vendors quote the two budgets separately (a latency number in one benchmark, a per-minute rate in a pricing PDF), but a production agent has to satisfy both on every turn, so this post puts them in one model and shows how Devotel Orbit exposes the trade-off before you deploy.
Why the two budgets belong on one row
A fast-but-expensive voice and a cheap-but-slow model both fail in production for opposite reasons. When the pipeline is metered once per turn, the useful question is "what did this component cost me in time and in money, given the choice I made?" Answering that per component is what turns tuning from a guessing game into a trade you can make deliberately.
The one-table budget
| Component | What moves the time | What moves the cost | Typical budget |
|---|---|---|---|
| Speech-to-text (STT) | Streaming vs batch, endpointing sensitivity, language | Per streamed audio-minute; custom vocabulary and language tiers | ~150–400 ms per turn; ~$0.005–$0.02 per audio-minute |
| Language model (LLM) | Model tier, prompt size, cache warmth, tool calls | Tokens in and out per turn — per-1M-token pricing | ~300–1500 ms to first token; the most variable line on the bill |
| Text-to-speech (TTS) | Streaming first byte, voice tier, failover warmth | Per character or per synthesized minute; standard vs cloned voices | ~100–300 ms to first byte; ~$0.01–$0.30 per minute |
Read the table as a joint constraint, not two independent ones. A cloned premium voice can cost twenty times a standard one; a frontier reasoning model can take five times as long as a routing-tier classifier. Neither is wrong — the mistake is paying for the premium tier on turns that do not need it.
STT: endpointing is a time-over-quality trade
The speech-to-text stage contributes two numbers: the transcription window itself, and the endpointing decision that decides when the caller has finished. Because every downstream stage waits on the endpoint, a conservative silence window adds its full duration to every turn. Shorter windows cut the time but raise the chance of interrupting the caller mid-thought; semantic endpointing that reads the transcript beat a blunt timeout is the way out. On cost, STT is the most stable line of the three — a per-minute streaming rate that barely moves unless you need a custom vocabulary or a long-tail language.
LLM: the line that swings both budgets at once
The language model is the only component where the time and the cost move together by an order of magnitude. A long system prompt re-read every turn, a cold prompt cache, or a tool call that re-injects a large context will push both the time-to-first-token and the token bill up at once; prompt caching and routing simple intents to a smaller model pull both back down. That is why the latency guide treats prompt caching and model tiering as latency techniques while the pricing guide lists them as cost controls — in this pipeline, they are the same knob.
TTS: stream the first byte, price the voice
Text-to-speech time is really time-to-first-byte: the agent starts speaking as the stream arrives, so a synthesiser built for streaming beats a nominally faster batch engine that buffers the whole reply. Keep a second provider warmed as failover so a slow first byte from one vendor does not stall the turn. Cost, meanwhile, is almost entirely a voice-tier decision — standard voices price by the character or minute, while cloned and premium voices price at a multiple of that. Pick the cloned voice where the brand calls for it and a standard voice everywhere else.
Seeing the trade before you deploy
Devotel Orbit ships four model presets — Balanced, High Intelligence, Ultra Fast, and Cost Saver — each annotated with its measured p50 latency, cost per minute, and a 1–5 quality score, so a team picks a published operating point instead of guessing in production. The span across the four presets is roughly 450 ms (Ultra Fast) to 1,400 ms (High Intelligence) at p50, with cost per minute following the same ordering ($0.03/min on Cost Saver, $0.22/min on the frontier model). Per-agent model tiering then lets you route the high-frequency simple intents to the fast tier and reserve the frontier model for the turns that genuinely need it, which is how both budgets come down together instead of one at the other's expense.
Reading the two budgets end to end
Sum the time column and you have the turn's time-to-first-audio, the number the caller actually judges; sum the cost column and you have the all-in per-minute the finance team actually pays. The time-to-first-audio target lives in the low one-second range, and the pricing guide's all-in rate lands in a wide band for exactly the reason above: the component choices you made. The human-agent TCO comparison then weighs that all-in figure against a fully-loaded human talk-minute, which is the comparison a budget owner is really running.
Frequently asked questions
Which component should I tune first?
The LLM. It dominates both budgets on most configurations, and the two highest-yield techniques — prompt caching and tiering high-frequency intents to a smaller model — cut time and cost on the same change.
Why is my TTS bill higher than the headline rate?
Voice tier. Cloned and premium voices price at a multiple of standard voices, so check the per-voice rate before you assume the provider's headline figure.
Is a sub-second turn worth the more expensive model?
Usually no. Consistency matters more than the absolute number: a stable sub-second p50 with a tight p95 beats a faster average with a fat tail. Spend the premium tier only on the turns that need it.